The Human Bottleneck: Why AI Adoption Stalls, and How to Fix It

Piles of pilots, and a much shorter list in production

Vinit opened by naming the two stories running in the press, one saying AI is running the world and one saying the infrastructure spend is a bubble. His view from inside enterprises is neither. What he sees is a large pile of ambitious pilots sitting next to a relatively small set of true production implementations. He was blunt about where the constraint lives, saying the model is fine and the difficulty is in whether enterprises can cope with the pace of innovation around them. Three years ago companies were shipping chatbots and voice agents. Within a year the conversation moved to agents running complex autonomous workflows, and organizational change simply does not move at that speed.

Business, data, and people, in that order

Asked what separates the companies doing this well, Vinit named three areas. The most common trap is treating AI as a technical challenge, which produces executives holding an impressive technology in search of somewhere to apply it. The leaders he calls visionaries flip the order and start from the business question, then work back to where AI produces something meaningful. He mentioned executives who put IA before AI, meaning information architecture before artificial intelligence. Second is data, since AI is only as good as what feeds it, and the companies moving fast have a unified platform that scales and extends across models as they leapfrog each other. Third is people and culture, where he was firm that technology produces no value on its own.

The center of excellence and the tiger team

A year ago Vinit spent two days with 40 CIOs in Washington, and the idea that came up most was the AI center of excellence. One large enterprise assembled a tiger team with representation from HR, marketing, finance, engineering, and product, and gave it a charter to build AI fluency across the company. That team brought in consultants to help rewrite job descriptions into roles aligned with how the work now gets done, pushed expectations down from the CEO in company wide email, and pushed enablement up from the floor. Vinit described a manufacturing CIO who asked him to bring that same fluency to factory floors in Guadalajara and Chennai, where workers with an iPad could build low code agents to automate their own work.

The third line item nobody budgeted for

Austen framed the shift plainly. Three years ago a tech company had two real expenses, people and cloud, and now inference has arrived as a third. Vinit hears the ROI question from boards constantly and was candid that nobody has solved it yet, including him. He described a CIO at one of the largest companies in the world admitting that AI sits in his budget as a cost with no ROI attached to it. The questions worth asking, in his framing, are who gains from the return, how AI produces it, and where it shows up. The leaders handling this best do three things, meaning they decide fast, they are maniacal about AI fluency, and they embed AI into core business processes rather than bolting it onto the edges.

Tokenomics, and giving Ferraris to your kids

Vinit has watched AI spend run out of control inside large organizations, and he named the terms circulating for it, including AI sprawl, agent sprawl, and token maxing. Engineering was the first place spend took off, since coding models produced real results and teams reported 60, 70, even 80 percent of code being written by AI. His counsel now is about fit. He used the image of handing a Ferrari to your kids to describe defaulting to the most expensive model for everything, and his teams walk customers through hybrid architectures where cheaper models carry most of the work. Austen offered the mirror image from the training side, where one employee spending a million a year on inference can be worth it while another uses a top tier model to check email.

The forward deployed engineer, and the unlearning it takes

Austen noted that mentions of forward deployed engineers have gone up roughly tenfold in six months. Vinit drew the line simply, saying engineers ship code and FDEs ship solutions. He told his own story to explain the appeal. His first job out of grad school had him writing modules in governance risk compliance with no idea how any of it reached the banks using it. Professional services put him in front of customers, and the move into sales engineering forced him to unlearn a great deal, since he knew the product so well that he stopped listening. An FDE role sits between those poles, letting an engineer stay a builder while working hip to hip with a customer. He listed what leaders should hire for, meaning technical depth, product intuition, empathy for how the customer sees it, communication, and extreme ownership.

The closing question

An audience question asked whether the future of AI is a technology challenge or a leadership one, and Vinit chose leadership. His reasoning was that the technology is guaranteed to get better, cheaper, and more efficient, while an organization's ability to adapt is guaranteed by nothing. His takeaway, offered as the one thing to carry out of the hour, was to stop asking which model to use and start asking what business friction you are trying to remove.